Yeah, thank you for the nice introduction and welcome everyone. So yes, indeed, this

talk is at least in part about this Spinnaker 2 and Spinnaker 2 system at here and also

the which is called the spin cloud. But I have put a few remarks on neuromorphic hardware

in general upfront to give some perspective on this hardware system and also some outlook

in the end to see where I see some fruitful directions actually for future developments.

So but let's just jump into it. So if you have attended the seminar two weeks back,

you know already a bit about neuromorphies from Katie Schumann from more from the kind

of computational side of things. I want to take now here kind of a hardware perspective

on the whole thing. So I have a background in chip design. So I view everything a bit

from that angle. And yeah, neuromorphic, as you might know, is tasked with investigating

how we can take over knowledge that we know about the human brain and how it processes

information to apply in integrated circuits or in hardware systems, as we say, and also

learn a bit from building those systems as well. We can do this on many different levels.

So our brain is a tremendously complex organ. So we can really go from single elements like

ion channels, the synapses and neurons up to the whole brain. And actually there is

in the middle in between kind of the global scale of the brain or brain areas and the local scale

of individual cells, there's not so much knowledge actually, because we cannot observe on those

scales. So we can measure individual neurons and bunches of neurons. And there's a detailed

understanding of those. And we can thanks to like, for example, fMRI or EEG, we can measure on the

global scale. But in between there are several orders of magnitude that are difficult really

to catch in measurements. And then thus we have to somehow make up some models or so to

bridge between those orders of magnitude. So now this is just a neuroscience side,

actually. And now we want to kind of strip off maybe all the biological details and use this for

a technical application. So we have to make a choice essentially on which level we want to actually

kind of rebuild what the brain does or mimic what the brain does and try to make use of it.

And just to pick out two core principles that have been in the neuromorphic community for decades

actually, from its beginning, it is one observation that's really on the, at least in my sketch here,

on the lowest level on the ion channels, is the observation that essentially a single ion channel

or the behavior of an ion channel is quite similar to a single transistor when we operate it in

sub-threshold operation. So sub-threshold operation is actually something for me as a circuit designer,

it's a pretty uncommon operating range. Normally you would operate your transistors kind of

above the threshold, but there is really this intriguing similarity. So people have been

building a lot of neuromorphic systems using this sub-threshold circuit technique. So that's one thing.

And if we go one level up to the synapse level, then there is the observation that essentially the

synapse has computation and storage merged. So it means we have some properties of the synapse,

so the coupling between two neurons that governs the strength of the synapse or its kind of synaptic

weight, as we also say. And at the same time, it transports the information from one side to the

other. So it also does the computation. And this is in stark contrast to the typical computer

architecture that has a separation of memory and processing. For good reasons, of course,

but it's not what we observe in the brain. So those two principles are really, I would say,

a lot of arguments in the neuromorphic computing team. But now if we look at some available bigger

scale systems, okay, this is not complete, but at least those are systems that, in my regard,

have achieved a remarkable size. Then you see that many of those, or the majority of those, actually

don't use those principles, which is a bit surprising. I have put here two more. So you see

on the top left, you see the sub threshold that I just talked about. Then connected to this is

really analog computation. So you don't use digital circuits, but analog ones. Then we have

the merge weight storage and computation, as we see in synapses. And another thing is that

people postulate to use asynchronous circuits because they say, okay, a clock or a synchronity

is not what we see in the brain, and it's an unnecessary overhead. But now you observe that

many of those systems don't really make use of those principles. And this comes maybe in the